A method and system for fault prediction and health management of a battery swap cabinet

Abstract

The application relates to the technical field of equipment health management, and discloses a battery replacement cabinet fault prediction and health management method and system. The method comprises the following steps: acquiring an original temperature sequence, an original humidity sequence, a battery module number and a battery replacement cabinet position number, performing abnormal value correction to obtain a corrected temperature sequence and a corrected humidity sequence; performing abnormal fluctuation point screening according to the corrected temperature sequence and the corrected humidity sequence to generate an abnormality judgment record; removing interference according to the abnormality judgment record to obtain a refined feature set; performing abnormal risk analysis according to the refined feature set to obtain an overheating risk report; generating an intervention instruction according to the overheating risk report to obtain an initial execution scheme; driving operation is performed according to the initial execution scheme to obtain an equipment operation log; and performing reliability evaluation according to the equipment operation log to obtain a reliability evaluation conclusion. The method can realize early warning and health management of battery replacement cabinet faults.

Description

Technical Field

[0001] This invention relates to the field of equipment health management technology, and in particular to a method and system for fault prediction and health management of a battery swapping cabinet. Background Technology

[0002] Currently, battery swapping cabinets face challenges during long-term operation, such as complex and ever-changing environments and irreversible damage to internal battery modules due to repeated charging and discharging. Subtle performance changes are often difficult to detect in a timely manner, and failure to accurately capture and intervene may trigger chain reactions or even safety accidents. Therefore, a fault prediction and health management technology is needed that can sense the equipment's operating status in real time, accurately identify subtle changes, and predict potential risks to improve the reliability and safety of equipment operation and ensure the stable operation of battery swapping cabinets in diverse scenarios.

[0003] In one existing technology, sensors deployed inside the battery swapping cabinet collect real-time temperature and voltage data from the battery modules and upload the data to a monitoring platform. The monitoring platform presets upper temperature limits and lower voltage limits. When the collected temperature exceeds the preset upper limit or the voltage falls below the preset lower limit, the system triggers an alarm and notifies maintenance personnel to conduct on-site repairs. Simultaneously, the system performs routine inspections of the equipment at fixed intervals, recording operating parameters and generating operational reports. Maintenance personnel then perform subsequent repairs or component replacements based on the alarm information and inspection records. This existing technology, by setting fixed thresholds for anomaly alarms and periodic inspections, only triggers a response after a fault occurs, leading to delayed problem detection and an inability to predict potential risks.

[0004] Therefore, existing technologies cannot provide early warning and health management for battery swapping cabinet failures. Summary of the Invention

[0005] This invention provides a method and system for fault prediction and health management of battery swapping cabinets, so as to realize early warning and health management of battery swapping cabinet faults.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for fault prediction and health management of a battery swapping cabinet, comprising:

[0007] Obtain the original temperature sequence, original humidity sequence, battery module number, and battery swapping cabinet location number;

[0008] Based on the original temperature sequence and the original humidity sequence, outlier correction is performed to obtain the corrected temperature sequence and the corrected humidity sequence;

[0009] Based on the corrected temperature sequence and the corrected humidity sequence, abnormal fluctuation points are filtered to obtain abnormal temperature fluctuation points and abnormal humidity fluctuation points. Based on the abnormal temperature fluctuation points, the abnormal humidity fluctuation points, the battery module number and the battery swapping cabinet location number, data association is performed to generate an anomaly judgment record.

[0010] Based on the anomaly determination record, the corrected temperature sequence or the corrected humidity sequence is windowed to obtain a subtle change sequence, and interference is removed based on the subtle change sequence to obtain a refined feature set;

[0011] Based on the refined feature set, risk probability prediction is performed to obtain a risk probability matrix, and an abnormal risk analysis is performed based on the risk probability matrix to obtain an overheating risk report.

[0012] Based on the overheating risk report, the overheating risk area is locked to obtain the high-risk area, and intervention instructions are generated based on the high-risk area to obtain the initial execution plan;

[0013] Based on the initial execution plan, the plan parameters are modified to obtain a modified execution plan, and the device is driven to run according to the modified execution plan to obtain the device operation log;

[0014] Based on the device operation log, instructions are adjusted according to the state deviation to obtain the simulated operation results. A reliability assessment is then performed based on the simulated operation results or the device operation log to obtain a reliability assessment conclusion.

[0015] Secondly, the present invention provides a fault prediction and health management system for a battery swapping cabinet, comprising:

[0016] The data acquisition module is used to acquire the original temperature sequence, original humidity sequence, battery module number, and battery swapping cabinet location number;

[0017] The noise reduction and correction module is used to perform outlier correction based on the original temperature sequence and the original humidity sequence to obtain the corrected temperature sequence and the corrected humidity sequence.

[0018] The anomaly determination module is used to filter abnormal fluctuation points based on the corrected temperature sequence and the corrected humidity sequence to obtain abnormal temperature fluctuation points and abnormal humidity fluctuation points, and to perform data association based on the abnormal temperature fluctuation points, the abnormal humidity fluctuation points, the battery module number and the battery swapping cabinet location number to generate anomaly determination records.

[0019] The interference removal module is used to perform windowing on the corrected temperature sequence or the corrected humidity sequence based on the anomaly determination record to obtain a subtle change sequence, and to remove interference based on the subtle change sequence to obtain a refined feature set.

[0020] The risk analysis module is used to predict the risk probability based on the refined feature set, obtain a risk probability matrix, and perform abnormal risk analysis based on the risk probability matrix to obtain an overheating risk report.

[0021] The instruction generation module is used to lock the overheating risk area based on the overheating risk report, obtain the high-risk area, and generate intervention instructions based on the high-risk area to obtain the initial execution plan;

[0022] The drive operation module is used to modify the scheme parameters according to the initial execution scheme to obtain a modified execution scheme, and to drive the operation according to the modified execution scheme to obtain the device operation log;

[0023] The output module is used to adjust instructions based on the state deviation according to the device operation log, obtain the simulation operation results, and perform a reliability assessment based on the simulation operation results or the device operation log to obtain a reliability assessment conclusion.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) This invention cleans the original temperature and humidity data by sliding window alignment and outlier removal, and calculates the fluctuation rate of adjacent sampling points by residual sequence. Combined with preset threshold, abnormal fluctuation points are screened out. Since the residual sequence can reflect the small deviation between the actual measured value and the theoretical benchmark value, the fluctuation rate further quantifies the severity of the deviation, thereby achieving accurate capture of subtle changes in the operating status of the equipment. It solves the problem that the existing technology only relies on fixed threshold alarm and is difficult to detect potential hidden dangers, and realizes early warning of faults.

[0026] (2) This invention predicts the initial risk probability through a decision tree model and uses the deviation coefficient between the actual temperature data and the risk probability set to weight and correct the risk probability. It constructs a risk probability matrix by combining the spatial location of the battery module, and screens out abnormal risk points and classifies risk levels through risk thresholds. Since the deviation coefficient can be calibrated according to the real-time temperature, the risk probability matrix associates the risk value with the spatial location, thereby achieving accurate positioning and classification of overheating risk. It solves the problem that the existing technology cannot quantify the degree of risk and is difficult to locate the risk location, and improves the accuracy and operability of fault prediction.

[0027] (3) This invention locks high-risk areas through a directed heat conduction graph, dynamically adjusts the sampling frequency to obtain high-frequency data, generates an execution plan through conflict detection and sorting, drives the action of the actuator through environmental correction, generates an operation log after collecting feedback, and then performs iterative adjustments and reliability assessments based on state deviations through digital twin simulation. Since the directed heat conduction graph can predict the heat diffusion path, the digital twin simulation can verify the adjustment effect without actual execution, thus forming a closed-loop optimization; it solves the problems of passive response and lack of verification of adjustment plan in the existing technology, and realizes the initiative and adaptability of health management. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the fault prediction and health management method for the battery swapping cabinet provided in the first embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of the fault prediction and health management system for the battery swapping cabinet provided in the second embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figure 1 The first embodiment of the present invention provides a method for fault prediction and health management of a battery swapping cabinet, comprising the following steps:

[0032] S11, obtain the original temperature sequence, original humidity sequence, battery module number and battery swapping cabinet location number;

[0033] S12, perform outlier correction based on the original temperature sequence and the original humidity sequence to obtain the corrected temperature sequence and the corrected humidity sequence;

[0034] S13, based on the corrected temperature sequence and the corrected humidity sequence, abnormal fluctuation points are filtered to obtain abnormal temperature fluctuation points and abnormal humidity fluctuation points. Based on the abnormal temperature fluctuation points, the abnormal humidity fluctuation points, the battery module number and the battery swapping cabinet location number, data association is performed to generate an anomaly judgment record.

[0035] S14, based on the anomaly determination record, the corrected temperature sequence or the corrected humidity sequence is windowed to obtain a subtle change sequence, and interference is removed based on the subtle change sequence to obtain a refined feature set;

[0036] S15, based on the refined feature set, perform risk probability prediction to obtain a risk probability matrix, and perform abnormal risk analysis based on the risk probability matrix to obtain an overheating risk report;

[0037] S16, Based on the overheating risk report, the overheating risk area is locked to obtain the high-risk area, and an intervention command is generated based on the high-risk area to obtain the initial execution plan;

[0038] S17. Based on the initial execution plan, the plan parameters are modified to obtain a modified execution plan, and the device is driven to run according to the modified execution plan to obtain the device operation log.

[0039] S18, based on the device operation log, adjust the instructions based on the state deviation to obtain the simulated operation results, and perform a reliability assessment based on the simulated operation results or the device operation log to obtain a reliability assessment conclusion.

[0040] In step S11, the original temperature sequence, original humidity sequence, battery module number, and battery swapping cabinet location number are obtained.

[0041] Specifically, temperature and humidity sensors deployed in each battery compartment of the battery swapping cabinet continuously collect temperature and humidity values ​​at preset global sampling intervals. Each sensor is physically bound to a fixed battery swapping cabinet compartment number and battery module number upon installation. During data collection, the sensor generates a raw record containing the current temperature and humidity values ​​along with the compartment number and battery module number. Each record includes the compartment number, battery module number, temperature value, humidity value, and a collection timestamp. The system arranges the temperature values ​​from multiple raw records into a raw temperature sequence and the humidity values ​​into a raw humidity sequence according to the collection time sequence. The corresponding compartment number and battery module number are then extracted as static identification information for subsequent steps.

[0042] The acquisition of raw temperature and humidity sequences relies on a preset global sampling time interval. This interval is pre-set based on the usage frequency of the battery swapping cabinet and the sensor performance; for example, the sampling interval is set to 1 second to ensure continuous monitoring of temperature and humidity changes. The compartment number and battery module number are preset static identifiers, uniformly assigned by the system and fixed in the equipment configuration during the deployment of the battery swapping cabinet, requiring no dynamic acquisition or calculation.

[0043] In step S12, outlier correction is performed based on the original temperature sequence and the original humidity sequence to obtain a corrected temperature sequence and a corrected humidity sequence, including:

[0044] Based on the original temperature sequence and the original humidity sequence, timestamp alignment is performed through a preset sliding window to obtain an aligned temperature sequence and an aligned humidity sequence.

[0045] Based on the aligned temperature sequence and the aligned humidity sequence, the mean value within each sliding window is calculated, and sampling points that deviate from the mean value by more than a preset multiple are removed to obtain the denoised temperature sequence and the denoised humidity sequence.

[0046] The denoised temperatures in the denoised temperature sequence that exceed a preset temperature threshold are replaced with the normal temperatures at the preceding and following times to obtain the corrected temperature sequence.

[0047] The denoised humidity values ​​that exceed the preset humidity range in the denoised humidity sequence are replaced with the boundary values ​​of the humidity range to obtain the corrected humidity sequence.

[0048] Specifically, the original temperature and humidity sequences are obtained in step S11. The system presets a fixed-length sliding window, the length of which is determined by historical data statistics. The 95th percentile of the maximum time deviation between adjacent sampling points in the historical data is selected as the window length, for example, the window length is set to 5 sampling points. The system sequentially covers the original temperature and humidity sequences with the sliding window. Within each window, the timestamps of all sampling points within the window are aligned to the center time of the window to achieve timestamp alignment, resulting in aligned temperature and humidity sequences.

[0049] For aligned temperature and humidity sequences, the system calculates the arithmetic mean of all sampling points within each sliding window. The preset rejection factor is determined by historical data statistics, selecting the upper limit of the normal fluctuation range in historical data; for example, the rejection factor is set to three times the standard deviation. The system calculates the standard deviation of all sampling points within each sliding window. When the absolute deviation of a sampling point exceeds three times the standard deviation, that sampling point is identified as an outlier and removed. After removal, the remaining sampling points are rearranged in chronological order to obtain the denoised temperature and humidity sequences.

[0050] The denoised temperature sequence may still contain abnormal temperature values ​​that exceed the normal response range of the sensor. The system presets temperature thresholds, which are determined by statistical analysis of historical temperature data distribution. For example, the 99th percentile of temperature values ​​under normal operating conditions is selected as the upper limit of the temperature threshold, and the 1st percentile is selected as the lower limit. When a temperature value in the denoised temperature sequence exceeds the preset upper limit or falls below the preset lower limit, the system locates the position of the abnormal temperature value in the sequence, extracts the temperature values ​​at the time preceding and following that position, and if both values ​​are within the normal range, calculates the arithmetic mean of the two values ​​as a replacement value. This replacement value replaces the abnormal temperature value, resulting in a corrected temperature sequence.

[0051] The denoised humidity sequence may contain abnormal humidity values ​​that exceed the sensor's physical measurement range. The system presets a humidity range, with the lower limit determined by the sensor's physical measurement lower limit and the upper limit by the sensor's physical measurement upper limit; for example, the humidity range can be set to 0 to 100. When a humidity value in the denoised humidity sequence is lower than the lower limit of the humidity range, the system replaces that value with the lower limit value; when a humidity value is higher than the upper limit value, the system replaces that value with the upper limit value; values ​​within the range remain unchanged, resulting in a corrected humidity sequence.

[0052] In step S13, abnormal fluctuation points are filtered based on the corrected temperature sequence and the corrected humidity sequence to obtain abnormal temperature fluctuation points and abnormal humidity fluctuation points. Data association is then performed based on the abnormal temperature fluctuation points, the abnormal humidity fluctuation points, the battery module number, and the battery swapping cabinet location number to generate an anomaly determination record, including:

[0053] The difference between the corrected temperature sequence and the preset reference temperature is calculated to obtain the temperature residual sequence, and the difference between the corrected humidity sequence and the preset reference humidity is calculated to obtain the humidity residual sequence.

[0054] The difference between two adjacent temperature residuals in the temperature residual sequence is calculated and then divided by a preset sampling time interval to obtain the temperature fluctuation rate. Similarly, the difference between two adjacent humidity residuals in the humidity residual sequence is calculated and then divided by the sampling time interval to obtain the humidity fluctuation rate.

[0055] The time points in the corrected temperature sequence that correspond to the temperature fluctuation rate exceeding the preset temperature fluctuation threshold are selected to obtain the abnormal temperature fluctuation points;

[0056] The time points in the corrected humidity sequence where the rate of change of humidity fluctuation exceeds a preset humidity fluctuation threshold are selected to obtain the abnormal humidity fluctuation points.

[0057] Based on the abnormal temperature fluctuation points, abnormal humidity fluctuation points, the battery module number, and the battery swapping cabinet location number, data association is performed to generate an anomaly determination record.

[0058] Specifically, the corrected temperature sequence and corrected humidity sequence output in step S12 are obtained. The preset reference temperature is determined by statistical analysis of historical operating data, selecting the 50th percentile of the temperature value of the battery swapping cabinet under stable operating conditions as the reference temperature; for example, the median of the temperature at all normal operating times in the historical data. The preset reference humidity is determined by statistical analysis of historical operating data, selecting the 50th percentile of the humidity value of the battery swapping cabinet under stable operating conditions as the reference humidity; for example, the median of the humidity at all normal operating times in the historical data. The system subtracts the reference temperature from each temperature value in the corrected temperature sequence to obtain the temperature difference corresponding to each sampling point, and arranges all differences in chronological order to form a temperature residual sequence. Similarly, the system subtracts the reference humidity from each humidity value in the corrected humidity sequence to obtain the humidity difference corresponding to each sampling point, and arranges all differences in chronological order to form a humidity residual sequence.

[0059] The preset sampling time interval is determined jointly by the sensor hardware performance and the monitoring requirements of the battery swapping cabinet. This interval has been set as a globally uniform value in step S11, for example, 1 second. The system traverses the temperature residual sequence. For two adjacent sampling points in the sequence, it calculates the difference between the temperature residual at the next moment and the temperature residual at the previous moment. This difference is divided by the sampling time interval to obtain the rate of temperature change per unit time, which is used as the rate of temperature fluctuation. The system traverses the humidity residual sequence. For two adjacent sampling points in the sequence, it calculates the difference between the humidity residual at the next moment and the humidity residual at the previous moment. This difference is divided by the sampling time interval to obtain the rate of humidity change per unit time, which is used as the rate of humidity fluctuation.

[0060] The preset temperature fluctuation threshold is determined by statistical analysis of historical operating data, selecting the 99th percentile of the rate of temperature fluctuation change under normal operating conditions in the historical operating data as the temperature fluctuation threshold. The system compares the rate of temperature fluctuation change corresponding to each sampling point with the temperature fluctuation threshold. If the rate of temperature fluctuation change exceeds the temperature fluctuation threshold, the time node corresponding to that sampling point in the corrected temperature sequence is marked as a temperature anomaly fluctuation point. The preset humidity fluctuation threshold is also determined by statistical analysis of historical operating data, selecting the 99th percentile of the rate of humidity fluctuation change under normal operating conditions in the historical operating data as the humidity fluctuation threshold. The system compares the rate of humidity fluctuation change corresponding to each sampling point with the humidity fluctuation threshold. If the rate of humidity fluctuation change exceeds the humidity fluctuation threshold, the time node corresponding to that sampling point in the corrected humidity sequence is marked as a humidity anomaly fluctuation point.

[0061] The system acquires the battery module number and battery swapping cabinet location number collected and stored in step S11. It associates the time points of abnormal temperature fluctuations with the corresponding battery module number and battery swapping cabinet location number, and similarly associates the time points of abnormal humidity fluctuations with the corresponding battery module number and battery swapping cabinet location number, merging these to generate anomaly determination records. Each anomaly determination record includes the anomaly type, time point, battery module number, battery swapping cabinet location number, and the corresponding fluctuation rate value. This serves as the output of step S13, providing accurate spatiotemporal positioning information for subsequent extraction of subtle change features.

[0062] In step S14, based on the anomaly determination record, the corrected temperature sequence or the corrected humidity sequence is windowed to obtain a subtle change sequence, and interference is removed based on the subtle change sequence to obtain a refined feature set, including:

[0063] Based on the time nodes in the anomaly determination record, sampling points are extracted from the corresponding correction sequence through a preset truncation window to obtain a subtle change sequence;

[0064] Based on the sequence of subtle changes, a frequency domain vector is obtained by performing a frequency domain transformation using the Fast Fourier Transform algorithm.

[0065] The amplitude of the frequency component in the frequency domain vector that matches the preset noise frequency library is set to zero to obtain a clean spectrum;

[0066] When the pure spectrum meets the preset data integrity conditions, the refined feature set is obtained by inverse transformation reconstruction using the inverse fast Fourier transform algorithm based on the pure spectrum.

[0067] When the clean spectrum does not meet the data integrity condition, the size of the truncating window is adjusted and truncating is performed again until the clean spectrum meets the data integrity condition, thus obtaining a refined feature set.

[0068] The process of constructing data integrity conditions includes:

[0069] Based on the pre-stored historical clean spectrum and the pre-stored historical spectrum quality set, the continuous proportion of effective frequency components is extracted through frequency band continuity analysis, and a continuity threshold is set.

[0070] Based on the historical pure spectrum, the total energy value is calculated through energy distribution analysis, the energy distribution range of the normal spectrum is statistically analyzed, and a lower limit and an upper limit of energy are set.

[0071] Based on the continuity threshold, the lower energy limit, and the upper energy limit, the frequency band continuity condition and the energy distribution condition are constructed.

[0072] By logically combining the frequency band continuity condition and the energy distribution condition, a condition is defined as data integrity when both conditions are met simultaneously, thus obtaining the data integrity condition.

[0073] The data integrity condition is verified by using historical spectrum samples. The accuracy rate and the false positive rate are calculated. When the false positive rate exceeds the preset false positive tolerance range, the continuity threshold, the lower limit of energy, or the upper limit of energy are adjusted. The process is iteratively optimized until the false positive rate does not exceed the false positive tolerance range, thus obtaining the final data integrity condition.

[0074] Specifically, the system acquires the anomaly determination record generated in step S13 and reads the corrected temperature and humidity sequences output in step S12. The anomaly determination record contains the anomaly type (temperature anomaly or humidity anomaly), time node, battery module number, and battery swapping cabinet location number. Based on the anomaly type, the system selects the corresponding corrected sequence. If the anomaly type is temperature anomaly, data is extracted from the corrected temperature sequence; if it is humidity anomaly, data is extracted from the corrected humidity sequence. The preset initial length of the extraction window is determined by statistical analysis of the time span of historical anomaly events. For example, the 95th percentile of the fluctuation duration in historical anomaly events is selected as the window radius, and twice this radius is used as the total window length. Using the time node in the anomaly determination record as the center, the system extracts continuous sampling points from the corresponding corrected sequence by counting the number of sampling points corresponding to the radius of each extraction window before and after it. These sampling points are then arranged in chronological order as a sequence of subtle changes.

[0075] The system performs a Fast Fourier Transform (FFT) on sequences of subtle changes. Specifically, the sequence is treated as a discrete-time signal, and the FFT algorithm is used to transform it into the frequency domain, outputting a complex sequence. Each complex number corresponds to a frequency component, with its magnitude being the amplitude of that frequency component and its phase being the phase angle. The amplitude value of this complex sequence is extracted and combined with the corresponding frequency values ​​to form a frequency vector, where each element contains both the frequency value and the amplitude value.

[0076] The pre-defined noise frequency library is constructed using historical data. Sensor data from the battery swapping cabinet is collected during normal operation and without abnormal events. A Fast Fourier Transform (FFT) is performed on the data to extract the amplitude values ​​of all frequency components. The amplitude distribution of each frequency component across multiple samplings is statistically analyzed. Frequency components whose mean amplitude exceeds the background noise threshold and matches a known interference source (such as the charger's switching frequency) are recorded as noise frequencies. All these frequency values ​​and their allowable deviation ranges are stored in the noise frequency library. The system iterates through each frequency component in the frequency domain vector, comparing its value with the noise frequencies in the library. If a match is found, the amplitude value of that frequency component is set to zero; otherwise, the original amplitude value is retained. After processing, a clean spectrum is obtained, with the same number of frequency components as the frequency domain vector, but the amplitude of the frequency components matching the noise is zero.

[0077] The system needs to determine whether the clean spectrum meets preset data integrity conditions. The process of constructing these conditions is as follows: A large number of historical clean spectrum samples are pre-collected, and each sample is labeled with a historical spectrum quality tag. These tags are automatically generated by the system based on physical laws (e.g., comparing the error between the original and reconstructed signals, checking the continuity of the spectrum shape, etc.). For each historical clean spectrum, a frequency band continuity analysis is performed. All frequency components are sorted from smallest to largest, and the number of continuous frequency segments with amplitudes greater than zero is calculated. The proportion of the longest continuous segment to all frequency components is taken as the continuity percentage. The continuity percentage of all historical complete samples is calculated, and the 10th percentile is taken as the continuity threshold. Simultaneously, an energy distribution analysis is performed. The total energy of each historical clean spectrum is calculated, which is the sum of the squares of the amplitudes of all frequency components. The total energy of all historical complete samples is calculated, and the 5th percentile is taken as the lower energy limit, and the 95th percentile as the upper energy limit. Based on the continuity threshold, a frequency band continuity condition is constructed, requiring that the continuity percentage of the current clean spectrum is greater than the continuity threshold. Based on the lower and upper energy limits, an energy distribution condition is constructed, requiring that the total energy of the current clean spectrum be between the lower and upper energy limits. The two conditions are logically combined to define data integrity as a condition that is met simultaneously, thus obtaining the initial data integrity condition. Then, this condition is validated using historical spectrum samples, and the accuracy and false positive rate are calculated. If the false positive rate exceeds the preset tolerance range, the continuity threshold, lower energy limit, or upper energy limit are adjusted. For example, the continuity threshold is adjusted in steps of 0.01, and the upper and lower energy limits are adjusted in steps of 5%, and the false positive rate is recalculated until it does not exceed the tolerance range, thus obtaining the final data integrity condition.

[0078] It should be further explained that the misjudgment tolerance range is determined based on the misjudgment rate distribution of normal spectrum in historical spectrum samples. Specifically, the 90th percentile of the misjudgment rate of all normal spectrum samples under the condition of complete initial data is taken as the upper limit of the tolerance range. For example, the tolerance range is set to a misjudgment rate of no more than five percent.

[0079] In step S14, the system uses the final data integrity condition to determine the clean spectrum. If the clean spectrum simultaneously satisfies the frequency band continuity condition and the energy distribution condition, it is determined to meet the data integrity condition. At this time, the system performs an inverse fast Fourier transform on the clean spectrum to convert the frequency domain signal back to the time domain, obtaining a reconstructed time domain sequence, which is the refined feature set. If the clean spectrum does not meet the data integrity condition, the system adjusts the size of the truncation window, for example, by increasing the window radius by a preset step size, such as adding 5 sampling points, and re-truncates sampling points from the corresponding corrected sequence. The fast Fourier transform, noise filtering, and data integrity condition determination are performed again, and this process is repeated until the clean spectrum meets the data integrity condition, and the corresponding refined feature set is output. If the window radius adjustment exceeds the preset maximum allowable range and still cannot meet the condition, the processing of the anomaly determination record is terminated, and it is marked as invalid; wherein, the maximum allowable range is a window radius of no more than one hundred sampling points, and this value is determined by the 99th percentile of the maximum duration of temperature fluctuations in historical anomaly events.

[0080] Through the aforementioned adaptive windowing and frequency domain noise filtering, the system can remove environmental interference from the original signal and extract subtle variation features that truly reflect the characteristics of the battery module itself. The refined feature set, as the output of step S14, provides high-quality input data for subsequent risk probability prediction, ensuring that risk analysis is based on pure and reliable signal features.

[0081] In step S15, risk probability prediction is performed based on the refined feature set to obtain a risk probability matrix, and anomaly risk analysis is performed based on the risk probability matrix to obtain an overheating risk report, including:

[0082] Based on the refined feature set, risk probability is predicted using a pre-built decision tree prediction model to obtain a risk probability set.

[0083] Based on the risk probability set and the corrected temperature sequence, the coefficients are solved using a pre-constructed deviation mapping function to obtain the deviation coefficients;

[0084] The risk probability matrix is ​​obtained by weighting and correcting the risk probability set based on the deviation coefficient, the battery module number, and the battery swapping cabinet location number.

[0085] The positions in the risk probability matrix that exceed a preset risk probability threshold are selected to obtain abnormal risk points;

[0086] Based on the abnormal risk points, the risk level is classified according to the pre-constructed risk level rules to obtain the risk level level, and an overheating risk report is generated based on the abnormal risk points and the risk level level.

[0087] The process of constructing the decision tree prediction model includes:

[0088] Based on the pre-stored historical refined feature set and the pre-stored historical risk probability set, an initial decision tree model is constructed using the decision tree algorithm. Information gain or Gini coefficient is used as the feature selection criterion, and the maximum depth and the minimum number of samples per leaf node are set.

[0089] Based on the historical refined feature set, the feature space is split by a recursive partitioning method. At each node, the splitting threshold of each feature is calculated, and the feature that reduces impurity the most is selected to branch with the threshold.

[0090] Based on the historical risk probability set, the maximum depth and the minimum number of samples per leaf node are optimized using a grid search method to determine the error tolerance range of the pruning strategy;

[0091] The generated decision tree is then post-pruned, and the prediction errors and error distribution sets before and after pruning are calculated.

[0092] When the error distribution set exceeds the error tolerance range, a cost-complexity pruning method is used to readjust the pruning strategy and iteratively update the tree structure until the error distribution set does not exceed the error tolerance range.

[0093] The splitting contribution of each feature in the decision tree is extracted by feature importance assessment, and the splitting contribution is correlated with the risk probability to construct the final decision tree prediction model.

[0094] Specifically, the system acquires the refined feature set output in step S14, the corrected temperature sequence output in step S12, and the battery module number and battery swapping cabinet location number collected in step S11. The refined feature set is a time-domain sequence reflecting the pure dynamic characteristics of the battery module within the abnormal fluctuation time window. The system inputs this refined feature set into a pre-built decision tree prediction model. Based on the mapping relationship between features and risk probabilities learned during training, the model outputs the risk probability value corresponding to each sampling point. The risk probability values ​​of all sampling points are arranged in chronological order to form a risk probability set.

[0095] The construction process of the decision tree prediction model is as follows: A large number of historical refined feature set samples are collected in advance, and each sample is labeled with a historical risk probability set, which is obtained by post-event statistical analysis of the failure rate during actual operation. The system uses the decision tree algorithm to construct the initial model, taking the historical refined feature set as the input feature and the historical risk probability set as the target output. During model training, information gain or Gini coefficient is used as the feature selection criterion. For each node, the system calculates the initial impurity of the current node's sample set, traverses all features and their possible splitting thresholds, and divides the samples into left and right subsets based on whether the feature value is greater than the threshold. The weighted sum of the impurities of the left and right subsets after splitting is calculated, and the feature and threshold that maximize the reduction of impurity are selected as the splitting conditions for that node. The system sets the maximum depth and the minimum number of samples per leaf node as pre-pruning parameters, for example, the maximum depth is set to 10 layers and the minimum number of samples per leaf node is set to 5 to prevent overfitting.

[0096] The training data for the decision tree prediction model is generated by collecting complete raw sensor data sequences and corresponding post-confirmed fault event records over a historical period, including the specific battery module, location, and precise time of the fault. Using each post-confirmed fault occurrence time as the center, the model traces back, simulating outlier correction, fluctuation point filtering, window truncation, and interference removal steps to generate a historical refined feature set of positive samples corresponding to the fault event. Simultaneously, during normal operation periods without faults, random time points are selected, and negative samples are generated through the same complete process. Each sample's historical risk probability set label is assigned based on whether it corresponds to a fault event; for example, samples before a fault are labeled with a high-risk probability of 1.0, normal samples with a low-risk probability of 0.1, and intermediate states are linearly interpolated over time. This constructs a training sample set completely consistent with the online inference process.

[0097] The system optimizes the maximum depth and minimum number of leaf nodes using a grid search method. Historical data is divided into training and validation sets. Predefined candidate depth values ​​(e.g., 5 to 20) and candidate minimum number of leaf nodes values ​​(e.g., 2 to 20) are traversed to train the model. Prediction errors (mean squared errors) are calculated on the validation set. The parameter combination that minimizes the validation error is selected as the final parameter, and the error distribution set (prediction error for each sample on the validation set) under these parameters is recorded. The error tolerance range of the pruning strategy is set based on the percentile of the validation error; for example, the 90th percentile of the validation error is used as the upper tolerance limit. Post-pruning is performed on the generated decision tree using a cost-complexity pruning method. The change in prediction error before and after pruning each subtree is calculated. If the error increment after pruning does not exceed the upper tolerance limit, pruning is performed. The tree structure is iteratively updated until the error distribution set does not exceed the tolerance range. Finally, the splitting contribution of each feature in the decision tree is extracted through feature importance assessment, that is, the total reduction in impurity brought about by each feature as a splitting node. The splitting contribution is correlated with the risk probability to determine the relationship between features and risk probability, and the final decision tree prediction model is constructed.

[0098] In step S15, the system uses the constructed decision tree prediction model to predict the refined feature set. For each sampling point in the refined feature set, the model judges layer by layer according to the branching conditions of the tree, and finally falls on a certain leaf node. The risk probability prediction value stored in the leaf node (i.e. the average risk probability of the training samples covered by the node) is the output of the sampling point. All outputs are arranged in order to form a risk probability set.

[0099] The system acquires a risk probability set and a corrected temperature sequence. A pre-defined deviation mapping function is determined through historical data regression analysis. A large number of historical samples are collected, including the risk probability set and the corresponding actual temperature sequence. The statistical relationship between the actual temperature and the model-predicted risk probability at each sampling point is calculated. Linear regression or piecewise linear fitting is used to obtain the mapping relationship from risk probability to expected temperature. The input to the deviation mapping function is the risk probability value, and the output is the corresponding expected temperature. The system iterates through each risk probability value in the risk probability set, calculates the expected temperature corresponding to that risk probability using the deviation mapping function, and simultaneously reads the actual temperature of that sampling point from the corrected temperature sequence, calculates the ratio of the actual temperature to the expected temperature, and uses this ratio as the deviation coefficient.

[0100] The training data for the deviation mapping function comes from the risk probability set and the corresponding actual corrected temperature sequence in the historical operation data. Within the window corresponding to each anomaly judgment record, the risk probability value output by the decision tree model is paired with the actual temperature value at the same time to form thousands of training sample pairs. The mapping relationship is obtained by fitting the data using the least squares method.

[0101] The system performs point-by-point multiplication of the deviation coefficient and the risk probability set to obtain the corrected risk probability value. Simultaneously, it acquires the battery module number and battery swapping cabinet location number corresponding to the sampling point. The corrected risk probability value is then combined with the location number, battery module number, and timestamp to form a structured risk probability matrix. The rows of this matrix correspond to different locations and battery modules, the columns correspond to time series, and the matrix elements are the corrected risk probability values.

[0102] The preset risk probability threshold is determined by statistical analysis of historical failure data. It collects the times when overheating failures occurred in the past, extracts the risk probability values ​​for a period of time prior to the failure, and takes the 10th percentile of these risk probability values ​​as the threshold. The system iterates through all elements in the risk probability matrix, comparing the risk probability value of each element with the risk probability threshold, and filters out matrix positions where the risk probability exceeds the threshold. Each position corresponds to an abnormal risk point, which includes the position number, battery module number, and timestamp information.

[0103] The system uses pre-built risk level rules to classify the risk level of abnormal risk points. These rules are determined through statistical analysis of historical failure samples. Historical overheating failure samples are collected, and the severity of the failure corresponding to different risk probability intervals before the failure occurs is statistically analyzed. The risk probability is divided into multiple intervals, each corresponding to a risk level. For example, a risk probability above the threshold up to 0.6 is considered mild risk, 0.6 to 0.8 is moderate risk, and above 0.8 is severe risk. The system determines the risk level of an abnormal risk point based on the interval in which its risk probability value falls. Finally, the system summarizes all abnormal risk points and their corresponding risk level, position number, battery module number, and timestamp to generate a structured overheating risk report. The report includes the location of the risk, the risk level, the risk probability value, and the recommended intervention priority. This overheating risk report, as the output of step S15, provides accurate decision-making basis for subsequent thermal management interventions.

[0104] In step S16, based on the overheating risk report, the overheating risk area is locked to obtain the high-risk area, and an intervention command is generated based on the high-risk area to obtain an initial execution plan, including:

[0105] Based on the overheating risk report, the battery cell is used as a node, the heat conduction medium and physical connection relationship between the battery cells are used as directed edges, and the preset heat transfer rate and preset attenuation coefficient are used as the weights of the edges to construct a directed graph model and obtain a directed graph of heat conduction.

[0106] Based on the directed heat conduction graph and the overheating risk report, the overheating risk region is identified using a breadth-first search algorithm, resulting in high-risk regions.

[0107] The sampling time interval of the high-risk area is adjusted to a preset high-frequency sampling interval, and sampling is performed again to obtain a high-frequency sampling stream;

[0108] Based on the high-frequency sampling stream, intervention actions and action parameters are matched using a preset thermal management strategy library to generate an intervention instruction sequence. Then, instruction pairs with time overlap in the intervention instruction sequence are marked to obtain conflict instruction records.

[0109] When the conflict instruction record is not empty, according to the conflict instruction record, the instruction with lower priority in the instruction pair is removed, and the intervention instruction sequence is sorted according to the execution time to obtain the initial execution plan;

[0110] When the conflict instruction record is empty, the intervention instruction sequence is directly used as the initial execution plan.

[0111] Specifically, the system retrieves the overheating risk report output in step S15. This report includes the location information of abnormal risk points (battery module number and battery swapping cabinet location number) and the corresponding risk level. The system reads the physical structure configuration data of the battery swapping cabinet, which pre-stores the number, spatial coordinates, heat transfer medium type, and physical connection relationship of each cell within the battery pack. Each cell is defined as a node, with attributes including cell number, spatial coordinates, and current risk level. For two cells in direct physical contact or connected via a heat transfer medium, a directed edge is established between them, its direction determined by the heat conduction direction (from the high-temperature region to the low-temperature region). The preset heat transfer rate is determined by the physical properties of the heat transfer medium and the contact area, obtained through experimental calibration. The cells are heated in a laboratory environment, the temperature rise rate of adjacent cells is measured, and the heat transfer rate is calculated and used as the heat transfer rate. The preset attenuation coefficient is determined by the length of the heat conduction path and the thermal resistance of the medium, and is obtained through theoretical calculation. Based on the thickness, thermal conductivity, and cross-sectional area of ​​the medium along the heat conduction path, the proportion of heat loss during the transfer process is calculated, and this proportion is used as the attenuation coefficient. The system uses the heat transfer rate and the attenuation coefficient as two weights of the directed edges to complete the construction of the directed graph model, resulting in a directed heat conduction graph.

[0112] The system uses a directed graph of thermal conductivity and high-risk cells from overheating risk reports as input, and executes a breadth-first search algorithm to identify risk areas. The algorithm starts with the risky cells identified in the overheating risk report, marking them as the first layer of nodes in the high-risk area. Subsequently, the algorithm traverses all outgoing edges of this node, obtaining adjacent nodes, and calculates the effectiveness of heat transfer based on the edge's heat transfer rate weight and attenuation coefficient weight. If the heat transfer rate exceeds a preset lower limit and the attenuation coefficient is lower than a preset upper limit, then heat is determined to be effectively transferred to the adjacent node, and this adjacent node is included in the second layer of nodes in the high-risk area. This process continues, expanding outwards layer by layer until no new nodes are included in a certain layer, or the number of expanded layers reaches a preset maximum diffusion layer number. The maximum diffusion layer number is determined based on the maximum physical distance between the cells arranged within the battery pack and the thermal conductivity attenuation characteristics; specifically, it is the number of hops traversed from the risky cell to the farthest cell whose temperature rise does not exceed 2 degrees Celsius above the ambient temperature, for example, 3 layers. All included nodes constitute a high-risk area, which includes the risky cell and its surrounding cells that may be affected by its heat.

[0113] The lower limit of the heat transfer rate is 0.1 degrees Celsius per second, and the upper limit of the attenuation coefficient is 0.5. These values ​​are obtained by measuring the cell heating experiment in a laboratory environment. When the temperature rise rate of adjacent cells is less than 0.1 degrees Celsius per second, the heat transfer is considered negligible; when the attenuation coefficient is higher than 0.5, the heat loss is considered too large and there is no need to include it in the high-risk area. These values ​​can be configured according to different cell models and thermal conductive media types.

[0114] The preset sampling time interval has been set to a globally uniform value in step S11, for example, 1 second. The system acquires this global sampling time interval and the preset high-frequency sampling interval. The high-frequency sampling interval is determined statistically by the minimum response time of temperature sudden changes in historical abnormal events. For example, the larger of the minimum sampling period supported by the sensor hardware and one-tenth of the average time for the temperature to rise from the normal value to the alarm value in historical abnormal events is taken as the high-frequency sampling interval. For example, if the minimum sampling period of the sensor is 0.1 seconds and the average rise time is 5 seconds, then the high-frequency sampling interval is 0.5 seconds. The system adjusts the sensor sampling interval corresponding to the high-risk area from the global sampling time interval to the high-frequency sampling interval and triggers resampling. During the resampling process, the system continuously collects temperature, voltage, and current data in the high-risk area according to the high-frequency sampling interval, forming a high-frequency sampling stream. The time density of this sampling stream is higher than that of the global sampling interval, which can capture transient parameter changes.

[0115] The pre-defined thermal management strategy library is constructed using historical intervention records and equipment operation manuals. The library contains multiple mapping rules, each consisting of a trigger condition and an execution action. The trigger condition defines the parameter range in the high-frequency sampling stream, such as a temperature rise rate exceeding 1 degree per second or a voltage drop rate exceeding 0.5 volts per second. The execution action defines the corresponding intervention action and its parameters, such as starting the liquid cooling pump's rotation speed, adjusting the fan's duty cycle, or disconnecting a specific circuit's relay number. The system matches the real-time parameter values ​​in the high-frequency sampling stream with the trigger conditions in the strategy library. When a rule's trigger condition is met, a corresponding intervention command is generated, containing the action type, action parameters, and execution time (current moment). After traversing all sampling points in the high-frequency sampling stream, an intervention command sequence is generated. The system further traverses the intervention command sequence, comparing the execution time intervals of every two commands. If the execution time intervals of two commands overlap, the command pair is marked as having time overlap, and all marked command pairs are recorded as conflict command records.

[0116] When the conflict instruction record is empty, the system directly sorts the intervention instruction sequence according to execution time and outputs it as the initial execution plan. When the conflict instruction record is not empty, the system processes the conflict instructions according to preset priority rules. The priority rules are determined through fault impact analysis, assessing the impact of each intervention action on equipment safety after failure; the higher the impact, the higher the priority. For example, cutting off the relay has a higher priority than adjusting the fan speed, and adjusting the fan speed has a higher priority than starting the liquid cooling pump. The system iterates through each instruction pair in the conflict instruction record, compares the priorities of the two instructions, and removes the instruction with the lower priority. After processing all conflict instruction pairs, the system reorders the remaining intervention instructions according to execution time to obtain the initial execution plan. The initial execution plan, as the output of step S16, provides a clear set of control instructions for subsequent execution and feedback.

[0117] In step S17, the scheme parameters are corrected according to the initial execution scheme to obtain a corrected execution scheme, and the drive is run according to the corrected execution scheme to obtain the device operation log, including:

[0118] Specifically, the system obtains the initial execution plan output in step S16. The initial execution plan contains a series of intervention instructions, each including the action type (e.g., adjusting fan speed, starting the liquid cooling pump, disconnecting the relay), action parameters (e.g., speed value, valve opening degree, relay number), and execution time. Simultaneously, the system acquires the current external ambient temperature of the battery swapping cabinet, which is obtained in real-time by an ambient temperature sensor deployed outside the cabinet.

[0119] The system uses pre-built environmental correction rules to adjust the action parameters in the initial execution plan. The process of constructing these rules is as follows: A large amount of historical operational data is collected beforehand, including actual effect data when performing the same intervention action under different ambient temperatures. For each intervention action, the ambient temperature is divided into multiple intervals. Within each interval, the action parameter values ​​required to achieve the target rate of temperature decrease in the battery compartment are statistically analyzed. The median of the action parameters from all successful samples within that interval is taken as the baseline parameter value for that interval. Using ambient temperature as the independent variable and the baseline parameter value as the dependent variable, piecewise linear interpolation is used to construct the mapping relationship between ambient temperature and parameter correction coefficients. For example, when the ambient temperature is below 10 degrees Celsius, the correction coefficient is 0.8; when the ambient temperature is between 10 and 25 degrees Celsius, the correction coefficient is 1.0; and when the ambient temperature is above 25 degrees Celsius, the correction coefficient is 1.2. The system iterates through each instruction in the initial execution plan, obtains the correction coefficient corresponding to the current ambient temperature interval, multiplies the original action parameters in the instruction with the correction coefficient to obtain the corrected action parameters, and replaces the original parameters with the corrected parameters to form the corrected execution plan.

[0120] It should be noted that the correction coefficient in the environmental correction rule is obtained by collecting actual cooling effect data after performing the same intervention action under different ambient temperatures in historical operation. With ambient temperature as the independent variable and cooling rate as the dependent variable, the least squares method is used to fit the linear regression equation. The correction coefficient is the predicted value of the regression equation at the midpoint of the corresponding temperature range divided by the predicted value at the base temperature.

[0121] The system sends the instructions in the modified execution plan to the bottom-level controller of the battery swapping cabinet in chronological order. The bottom-level controller drives the corresponding actuators to perform actions, such as sending pulse width modulation signals to the fan to adjust its speed, sending switching signals to the liquid cooling pump to start the cooling cycle, and sending cut-off signals to the relay to disconnect the fault circuit. During execution, the system continuously collects operational feedback data through sensors deployed in the battery compartment, including physical quantities such as the actual fan speed, motor operating current, battery compartment temperature, voltage, and current. All collected feedback data are arranged in chronological order to form an operational feedback set.

[0122] The system extracts process parameters from the operational feedback set. The extraction of process parameters includes the following calculations: for response time, the system calculates the time difference between the moment the command is issued and the moment the actuator actually begins to move; for steady-state error, the system calculates the absolute difference between the actual action parameter after the actuator reaches a steady state and the command target parameter; for overshoot, the system calculates the difference between the maximum peak value of the actual parameter exceeding the target parameter during the adjustment process and the target parameter; for temperature change rate, the system calculates the ratio of the decrease in battery compartment temperature within a preset time window (e.g., 5 seconds) to the length of the time window; for peak current, the system extracts the maximum current value during motor startup. All the extracted process parameters are organized chronologically to form a status log set. Each status log record includes a timestamp, command number, process parameter type, and parameter value.

[0123] The system acquires a status log set containing status data across multiple dimensions, which may have different sampling periods. For example, the temperature sensor has a sampling period of 500 milliseconds, while the fan speed sensor has a sampling period of 100 milliseconds, and their timestamps do not completely overlap. The system uses a unified absolute time axis as a reference, dividing the time axis into equally spaced time nodes, with the interval between time nodes calculated as the least common divisor of the sampling periods of all sensors. For each time node, the system searches the status log set for the nearest sampling points before and after that node, and uses a linear interpolation method to calculate the corresponding status value for that node. Specifically, for temperature data, if there are temperature sampling points before and after a time node, the slope of the line connecting the two sampling points is calculated, and the interpolated temperature value is calculated based on the time difference from the time node to the previous sampling point. After interpolating the status data for all dimensions, a fully aligned log set with complete status values ​​for each dimension at each time node is obtained.

[0124] The system merges all status data at the same time point in the aligned log set. The merging operation combines multi-dimensional status values ​​such as temperature, rotational speed, current, and voltage at the same time point into a single record. This record includes a timestamp, bay number, battery module number, and the status values ​​for all dimensions at that moment. All records from all time points are arranged chronologically to form the equipment operation log. As the output of step S17, the equipment operation log fully records the entire process of actuator response and battery module status changes, providing a precise data foundation for subsequent stability assessment and reliability analysis.

[0125] In step S18, based on the device operation log, instruction adjustments are made according to the state deviation to obtain the simulated operation results. A reliability assessment is then performed based on the simulated operation results or the device operation log to obtain a reliability assessment conclusion, including:

[0126] Extract the timing waveforms from the device operation log, and perform scene matching on the timing waveforms using a preset scene feature library to obtain the operation scene category;

[0127] Based on the operational scenario category, a state deduction is performed using a pre-built stability analysis model to obtain an operational state sequence, and the deviation between the operational state sequence and a preset stable operation standard is calculated to obtain the state deviation.

[0128] When the state deviation exceeds the preset state deviation threshold, adjustment measures and adjustment parameters are matched according to the running state sequence through the preset running adjustment strategy library to generate adjustment instructions, and the adjustment instructions are input into the digital twin simulation environment to obtain the simulation running results;

[0129] Based on the simulation results, a reliability assessment is performed using pre-built reliability judgment rules to obtain a reliability assessment conclusion.

[0130] When the state deviation does not exceed the preset state deviation threshold, a reliability assessment is performed based on the device operation log and the reliability judgment rule to obtain a reliability assessment conclusion.

[0131] Specifically, the system obtains the device operation log output in step S17. The device operation log records the entire process of temperature, voltage, current, and actuator response parameters of the battery module changing over time after the intervention command is executed. The system extracts time-series waveforms from the device operation logs. Specifically, this involves extracting the values ​​of the same physical quantity (e.g., temperature) from the logs in chronological order to form a time-series sequence for that physical quantity. The system merges multiple time-series sequences, such as temperature, voltage, and current sequences, to form a multi-dimensional time-series waveform. A preset scenario feature library is constructed using historical operation data, collecting a large number of historical device operation logs and labeling each log with its operation scenario category (e.g., high-frequency pulse charging scenario, high-temperature standby scenario, low-temperature start-up scenario, etc.). For logs under each scenario category, the system extracts the statistical features of their time-series waveforms, including mean, variance, rise time, fall time, peak value, and frequency, and stores these feature vectors in the feature library. The system matches the feature vector of the current time-series waveform with the feature vectors of each scene category in the feature library, calculates the Euclidean distance or cosine similarity, and selects the scene category with the closest distance or the highest similarity as the current running scene category, thus obtaining the running scene category.

[0132] The system acquires the operating scenario category and invokes a pre-built stability analysis model. The stability analysis model is built through training on historical data, collecting historical device operation logs and their corresponding operating scenario categories. For each log entry, an operating state sequence label is extracted, labeled according to physical laws, such as charging full load period, rapid temperature rise period, thermal equilibrium period, and cooling period. The model employs a recurrent neural network structure. The input layer receives time-series waveform features, the hidden layer uses long short-term memory units to capture temporal dependencies, and the output layer outputs predicted values ​​of the state sequence within a preset future time window. During model training, the input is the time-series waveform features of historical logs, and the output is the corresponding state sequence label. The network weights are optimized using a backpropagation algorithm. After model training, the system inputs the time-series waveform features corresponding to the current operating scenario category into the model. After forward computation, the model outputs an operating state sequence, which describes the various stages the device will experience in the future. Simultaneously, the model calculates the state deviation value by extracting the real-time parameters of the current device operation (such as the current temperature rise rate) and subtracting them from the model's output safety baseline state, such as the safe temperature rise rate. This difference is used as the state deviation.

[0133] The state sequence labels required for training the stability analysis model are generated through manual annotation combined with physical laws. Specifically, for each historical equipment operation log, stages such as full charging period, rapid temperature rise period, thermal equilibrium period, and cooling period are marked according to the temperature curve shape, and consecutive time points within the same stage are assigned the same state label to form the output sequence required for supervised learning.

[0134] The preset state deviation threshold is determined through statistical analysis of historical safe operation data. A large amount of historical data on equipment safe operation is collected, and the state deviation value for each run is calculated. The 95th percentile of these deviation values ​​is taken as the state deviation threshold. The system compares the current state deviation value with the state deviation threshold. If the state deviation exceeds the threshold, an adjustment branch is executed; otherwise, an evaluation branch is executed directly.

[0135] In the adjustment branch, the system calls a preset operation adjustment strategy library based on the operating state sequence. The operation adjustment strategy library is built using historical optimization records, collecting a large number of adjustment cases where state deviations exceeded thresholds during historical operations. It records the operating state sequence, adjustment measures taken, adjustment parameters, and the effects of each case. The strategy library is stored in rule form, with each rule containing a condition part (a specific pattern of the operating state sequence) and a conclusion part (adjustment measures and adjustment parameters). The system matches the current operating state sequence with the condition part in the strategy library, selects the rule with the highest matching degree, extracts the corresponding adjustment measures and adjustment parameters, and generates an adjustment instruction. The adjustment instruction includes the adjustment type (e.g., reducing charging power, limiting charging current, adjusting voltage threshold), adjustment parameters (e.g., reduction percentage, current upper limit), and execution timing. The system inputs the adjustment instruction into a digital twin simulation environment, which is a virtual mirror model of the battery swapping cabinet, including the physical characteristics, thermodynamic model, and control logic of the battery modules, capable of simulating the state changes of the equipment after the adjustment instruction is executed. After the simulation environment executes the adjustment command, it outputs the simulation results, which include simulation data such as the adjusted temperature field distribution, power load curve, and voltage and current change trends.

[0136] It is worth noting that the digital twin simulation environment is constructed based on the physical parameters and measured response data of the battery swapping cabinet. A thermodynamic differential equation model of the battery module is established, and the heat capacity and thermal resistance parameters are identified through a step response experiment. A power output model of the charger is established, and the efficiency curve is fitted through a load experiment. A simplified fluid dynamics model of the fan and liquid cooling pump is established, and the transfer function is determined through a speed-airflow calibration experiment. The above sub-models are integrated and calibrated in the simulation platform to ensure that the root mean square error between the simulation output and the measured data does not exceed five percent.

[0137] In the evaluation branch, the system directly uses the equipment operation log output in step S17 as the evaluation basis. The system calls the pre-built reliability judgment rules. The reliability judgment rules are determined through historical fault data statistics. A large amount of historical equipment operation data is collected, and the data is divided into safe operation range and fault operation range. The distribution range of key indicators (such as maximum temperature, temperature rise rate, and voltage fluctuation range) within the safe range is analyzed. The 1st percentile of each indicator is taken as the lower limit, and the 99th percentile is taken as the upper limit to form the judgment rules. The rule is defined as follows: when all key indicators are between the lower limit and the upper limit, it is judged as reliable; otherwise, it is judged as unreliable. The system compares the key indicators in the simulation operation results or equipment operation log one by one. If all indicators are within the safe range, the reliability assessment conclusion is reliable; if any indicator exceeds the safe range, the reliability assessment conclusion is unreliable, and the value and location of the indicator that exceeds the safe range are output. The reliability assessment conclusion is the output of step S18, providing the final basis for the operation and maintenance decision of the battery swapping cabinet.

[0138] Reference Figure 2 The second embodiment of the present invention provides a fault prediction and health management system for a battery swapping cabinet, comprising:

[0139] The data acquisition module is used to acquire the original temperature sequence, original humidity sequence, battery module number, and battery swapping cabinet location number;

[0140] The noise reduction and correction module is used to perform outlier correction based on the original temperature sequence and the original humidity sequence to obtain the corrected temperature sequence and the corrected humidity sequence.

[0141] The anomaly determination module is used to filter abnormal fluctuation points based on the fluctuation change rate according to the corrected temperature sequence and the corrected humidity sequence to obtain abnormal temperature fluctuation points and abnormal humidity fluctuation points, and to perform data association based on the abnormal temperature fluctuation points, the abnormal humidity fluctuation points, the battery module number and the battery swapping cabinet location number to generate anomaly determination records.

[0142] The interference removal module is used to perform windowing on the corrected temperature sequence or the corrected humidity sequence based on the anomaly determination record to obtain a subtle change sequence, and to remove interference based on the subtle change sequence to obtain a refined feature set.

[0143] The risk analysis module is used to predict the risk probability based on the refined feature set, obtain a risk probability matrix, and perform abnormal risk analysis based on the risk probability matrix to obtain an overheating risk report.

[0144] The instruction generation module is used to lock the overheating risk area based on the overheating risk report, obtain the high-risk area, and generate intervention instructions based on the high-risk area to obtain the initial execution plan;

[0145] The drive operation module is used to modify the scheme parameters according to the initial execution scheme to obtain a modified execution scheme, and to drive the operation according to the modified execution scheme to obtain the device operation log;

[0146] The output module is used to adjust instructions based on the state deviation according to the device operation log, obtain the simulation operation results, and perform a reliability assessment based on the simulation operation results or the device operation log to obtain a reliability assessment conclusion.

[0147] It should be noted that the fault prediction and health management system for a battery swapping cabinet provided in this embodiment of the invention is used to execute all the process steps of the fault prediction and health management method for a battery swapping cabinet in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0148] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0149] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for fault prediction and health management of a battery swapping cabinet, characterized in that, include: Obtain the original temperature sequence, original humidity sequence, battery module number, and battery swapping cabinet location number; Based on the original temperature sequence and the original humidity sequence, outlier correction is performed to obtain the corrected temperature sequence and the corrected humidity sequence; Based on the corrected temperature sequence and the corrected humidity sequence, abnormal fluctuation points are filtered to obtain abnormal temperature fluctuation points and abnormal humidity fluctuation points. Based on the abnormal temperature fluctuation points, the abnormal humidity fluctuation points, the battery module number and the battery swapping cabinet location number, data association is performed to generate an anomaly judgment record. Based on the anomaly determination record, the corrected temperature sequence or the corrected humidity sequence is windowed to obtain a subtle change sequence, and interference is removed based on the subtle change sequence to obtain a refined feature set; Based on the refined feature set, risk probability prediction is performed to obtain a risk probability matrix, and an abnormal risk analysis is performed based on the risk probability matrix to obtain an overheating risk report. Based on the overheating risk report, the overheating risk area is locked to obtain the high-risk area, and intervention instructions are generated based on the high-risk area to obtain the initial execution plan; Based on the initial execution plan, the plan parameters are modified to obtain a modified execution plan, and the device is driven to run according to the modified execution plan to obtain the device operation log; Based on the device operation log, instructions are adjusted according to the state deviation to obtain the simulated operation results. A reliability assessment is then performed based on the simulated operation results or the device operation log to obtain a reliability assessment conclusion.

2. The method for fault prediction and health management of the battery swapping cabinet according to claim 1, characterized in that, The step of performing outlier correction based on the original temperature sequence and the original humidity sequence to obtain corrected temperature and humidity sequences includes: Based on the original temperature sequence and the original humidity sequence, timestamp alignment is performed through a preset sliding window to obtain an aligned temperature sequence and an aligned humidity sequence. Based on the aligned temperature sequence and the aligned humidity sequence, the mean value within each sliding window is calculated, and sampling points that deviate from the mean value by more than a preset multiple are removed to obtain the denoised temperature sequence and the denoised humidity sequence. The denoised temperatures in the denoised temperature sequence that exceed a preset temperature threshold are replaced with the normal temperatures at the preceding and following times to obtain the corrected temperature sequence. The denoised humidity values ​​that exceed the preset humidity range in the denoised humidity sequence are replaced with the boundary values ​​of the humidity range to obtain the corrected humidity sequence.

3. The fault prediction and health management method for the battery swapping cabinet according to claim 1, characterized in that, The process involves filtering abnormal fluctuation points based on the corrected temperature sequence and the corrected humidity sequence to obtain abnormal temperature and humidity fluctuation points. Then, based on these abnormal temperature and humidity fluctuation points, the battery module number, and the battery swapping cabinet location number, data association is performed to generate an anomaly determination record, including: The difference between the corrected temperature sequence and the preset reference temperature is calculated to obtain the temperature residual sequence, and the difference between the corrected humidity sequence and the preset reference humidity is calculated to obtain the humidity residual sequence. The difference between two adjacent temperature residuals in the temperature residual sequence is calculated and then divided by a preset sampling time interval to obtain the temperature fluctuation rate. Similarly, the difference between two adjacent humidity residuals in the humidity residual sequence is calculated and then divided by the sampling time interval to obtain the humidity fluctuation rate. The time points in the corrected temperature sequence that correspond to the temperature fluctuation rate exceeding the preset temperature fluctuation threshold are selected to obtain the abnormal temperature fluctuation points; The time points in the corrected humidity sequence where the rate of change of humidity fluctuation exceeds a preset humidity fluctuation threshold are selected to obtain the abnormal humidity fluctuation points. Based on the abnormal temperature fluctuation points, abnormal humidity fluctuation points, the battery module number, and the battery swapping cabinet location number, data association is performed to generate an anomaly determination record.

4. The method for fault prediction and health management of the battery swapping cabinet according to claim 1, characterized in that, The step involves windowing the corrected temperature sequence or the corrected humidity sequence based on the anomaly determination record to obtain a subtle change sequence, and then removing interference from the subtle change sequence to obtain a refined feature set, including: Based on the time nodes in the anomaly determination record, sampling points are extracted from the corresponding correction sequence through a preset truncation window to obtain a subtle change sequence; Based on the sequence of subtle changes, a frequency domain vector is obtained by performing a frequency domain transformation using the Fast Fourier Transform algorithm. The amplitude of the frequency component in the frequency domain vector that matches the preset noise frequency library is set to zero to obtain a clean spectrum; When the pure spectrum meets the preset data integrity conditions, the refined feature set is obtained by inverse transformation reconstruction using the inverse fast Fourier transform algorithm based on the pure spectrum. When the clean spectrum does not meet the data integrity condition, the size of the truncating window is adjusted and truncating is performed again until the clean spectrum meets the data integrity condition, thus obtaining a refined feature set.

5. The method for fault prediction and health management of the battery swapping cabinet according to claim 4, characterized in that, The process of constructing the data integrity conditions includes: Based on the pre-stored historical clean spectrum and the pre-stored historical spectrum quality set, the continuous proportion of effective frequency components is extracted through frequency band continuity analysis, and a continuity threshold is set. Based on the historical pure spectrum, the total energy value is calculated through energy distribution analysis, the energy distribution range of the normal spectrum is statistically analyzed, and a lower limit and an upper limit of energy are set. Based on the continuity threshold, the lower energy limit, and the upper energy limit, the frequency band continuity condition and the energy distribution condition are constructed. By logically combining the frequency band continuity condition and the energy distribution condition, a condition is defined as data integrity when both conditions are met simultaneously, thus obtaining the data integrity condition. The data integrity condition is verified by using historical spectrum samples. The accuracy rate and the false positive rate are calculated. When the false positive rate exceeds the preset false positive tolerance range, the continuity threshold, the lower limit of energy, or the upper limit of energy are adjusted. The process is iteratively optimized until the false positive rate does not exceed the false positive tolerance range, thus obtaining the final data integrity condition.

6. The method for fault prediction and health management of the battery swapping cabinet according to claim 1, characterized in that, The process of predicting risk probability based on the refined feature set to obtain a risk probability matrix, and performing anomaly risk analysis based on the risk probability matrix to obtain an overheating risk report, includes: Based on the refined feature set, risk probability is predicted using a pre-built decision tree prediction model to obtain a risk probability set. Based on the risk probability set and the corrected temperature sequence, the coefficients are solved using a pre-constructed deviation mapping function to obtain the deviation coefficients; The risk probability matrix is ​​obtained by weighting and correcting the risk probability set based on the deviation coefficient, the battery module number, and the battery swapping cabinet location number. The positions in the risk probability matrix that exceed a preset risk probability threshold are selected to obtain abnormal risk points; Based on the abnormal risk points, the risk level is classified according to the pre-constructed risk level rules to obtain the risk level level, and an overheating risk report is generated based on the abnormal risk points and the risk level level.

7. The fault prediction and health management method for the battery swapping cabinet according to claim 6, characterized in that, The process of constructing the decision tree prediction model includes: Based on the pre-stored historical refined feature set and the pre-stored historical risk probability set, an initial decision tree model is constructed using the decision tree algorithm. Information gain or Gini coefficient is used as the feature selection criterion, and the maximum depth and the minimum number of samples per leaf node are set. Based on the historical refined feature set, the feature space is split by a recursive partitioning method. At each node, the splitting threshold of each feature is calculated, and the feature that reduces impurity the most is selected to branch with the threshold. Based on the historical risk probability set, the maximum depth and the minimum number of samples per leaf node are optimized using a grid search method to determine the error tolerance range of the pruning strategy; The generated decision tree is then post-pruned, and the prediction errors and error distribution sets before and after pruning are calculated. When the error distribution set exceeds the error tolerance range, a cost-complexity pruning method is used to readjust the pruning strategy and iteratively update the tree structure until the error distribution set does not exceed the error tolerance range. The splitting contribution of each feature in the decision tree is extracted by feature importance assessment, and the splitting contribution is correlated with the risk probability to construct the final decision tree prediction model.

8. The method for fault prediction and health management of the battery swapping cabinet according to claim 3, characterized in that, The process involves identifying overheating risk areas based on the overheating risk report, determining high-risk areas, and generating intervention instructions based on these high-risk areas to obtain an initial execution plan, including: Based on the overheating risk report, the battery cell is used as a node, the heat conduction medium and physical connection relationship between the battery cells are used as directed edges, and the preset heat transfer rate and preset attenuation coefficient are used as the weights of the edges to construct a directed graph model and obtain a directed graph of heat conduction. Based on the directed heat conduction graph and the overheating risk report, the overheating risk region is identified using a breadth-first search algorithm, resulting in high-risk regions. The sampling time interval of the high-risk area is adjusted to a preset high-frequency sampling interval, and sampling is performed again to obtain a high-frequency sampling stream; Based on the high-frequency sampling stream, intervention actions and action parameters are matched using a preset thermal management strategy library to generate an intervention instruction sequence. Then, instruction pairs with time overlap in the intervention instruction sequence are marked to obtain conflict instruction records. When the conflict instruction record is not empty, according to the conflict instruction record, the instruction with lower priority in the instruction pair is removed, and the intervention instruction sequence is sorted according to the execution time to obtain the initial execution plan; When the conflict instruction record is empty, the intervention instruction sequence is directly used as the initial execution plan.

9. The method for fault prediction and health management of the battery swapping cabinet according to claim 1, characterized in that, The step of adjusting instructions based on state deviation according to the device operation log to obtain simulated operation results, and performing a reliability assessment based on the simulated operation results or the device operation log to obtain a reliability assessment conclusion, includes: Extract the timing waveforms from the device operation log, and perform scene matching on the timing waveforms using a preset scene feature library to obtain the operation scene category; Based on the operational scenario category, a state deduction is performed using a pre-built stability analysis model to obtain an operational state sequence, and the deviation between the operational state sequence and a preset stable operation standard is calculated to obtain the state deviation. When the state deviation exceeds the preset state deviation threshold, adjustment measures and adjustment parameters are matched according to the running state sequence through the preset running adjustment strategy library to generate adjustment instructions, and the adjustment instructions are input into the digital twin simulation environment to obtain the simulation running results; Based on the simulation results, a reliability assessment is performed using pre-built reliability judgment rules to obtain a reliability assessment conclusion. When the state deviation does not exceed the preset state deviation threshold, a reliability assessment is performed based on the device operation log and the reliability judgment rule to obtain a reliability assessment conclusion.

10. A fault prediction and health management system for a battery swapping cabinet, characterized in that, include: The data acquisition module is used to acquire the original temperature sequence, original humidity sequence, battery module number, and battery swapping cabinet location number; The noise reduction and correction module is used to perform outlier correction based on the original temperature sequence and the original humidity sequence to obtain the corrected temperature sequence and the corrected humidity sequence. The anomaly determination module is used to filter abnormal fluctuation points based on the corrected temperature sequence and the corrected humidity sequence to obtain abnormal temperature fluctuation points and abnormal humidity fluctuation points, and to perform data association based on the abnormal temperature fluctuation points, the abnormal humidity fluctuation points, the battery module number and the battery swapping cabinet location number to generate anomaly determination records. The interference removal module is used to perform windowing on the corrected temperature sequence or the corrected humidity sequence based on the anomaly determination record to obtain a subtle change sequence, and to remove interference based on the subtle change sequence to obtain a refined feature set. The risk analysis module is used to predict the risk probability based on the refined feature set, obtain a risk probability matrix, and perform abnormal risk analysis based on the risk probability matrix to obtain an overheating risk report. The instruction generation module is used to lock the overheating risk area based on the overheating risk report, obtain the high-risk area, and generate intervention instructions based on the high-risk area to obtain the initial execution plan; The drive operation module is used to modify the scheme parameters according to the initial execution scheme to obtain a modified execution scheme, and to drive the operation according to the modified execution scheme to obtain the device operation log; The output module is used to adjust instructions based on the state deviation according to the device operation log, obtain the simulation operation results, and perform a reliability assessment based on the simulation operation results or the device operation log to obtain a reliability assessment conclusion.

Images